多阶段的分割算法¶
- Intensity $I = \frac{R + G + B}{3}$ Saturation $S = 1 - \frac{\min(R,G,B)}{I} $ $$ I > 0.4 * I_{max}\ \ and\ \ S < 0.6 * S_{max} $$
- gradient magnitude $ |\triangledown G| = \sqrt{(\frac{\partial I}{\partial x})^2 + (\frac{\partial I}{\partial y})^2}$ $$ |\triangledown G| > 0.15 $$
- Gaussin Mixture Model cluster
概率密度函数
$$
p(x_{i}) = \sum_{k=1}^{K}\pi_{k}P(x_i|j) \\
= \sum_{k=1}^{K}\pi_{k} \mathcal{N}(x_i|\mu_k,\Sigma_k)
$$
- S -- V feature space, where S and V are Saturation and Value of the standard HSV color space
- 似然函数： $$ \prod_{i=1}^{N}p(x_i) $$
- log似然函数： $$ \sum_{i=1}^{N}\log p(x_i) = \sum_{i=1}^{N}\log (\sum_{k=1}^{K}\pi_{k} \mathcal{N}(x_i|\mu_k,\Sigma_k)) $$
- EM算法：引入因变量z，似然函数变为：$$\prod_{i=1}^{N}p(x_i,z_i) = \prod_{i=1}^{N}p(z_i)p(x_i|z_i) = \prod_{i=1}^{N} \prod_{k=1}^{K}(\pi_k)^{z_k}\mathcal{N}(x_i|\mu_k,\Sigma_k)^{z_k}$$
- log似然：$$\sum_{i=1}^{N}\sum_{i=1}^{N}z_{ik}\{\log \pi_k + \log \mathcal{N}(x_i|\mu_k,\Sigma_k)\}$$
- 估计$z_k$,根据后验概率取期望：$$z_k = 0 * p(z_k = 0|x) + 1 * p(z_k = 1 | x) \\ = p(z_k = 1 | x) \\= \frac{p(x)p(x|z_k =1)}{p(z_k)} \\= \frac{\pi_k\mathcal{N}(x|\mu_k,\Sigma_k)}{\sum_{j=1}^{K}\pi_j\mathcal{N}(x|\mu_j,\Sigma_j)}$$
填充算法¶
In [10]:
plt.figure(figsize=(40,40))
plt.subplot(1,2,1)
plt.imshow(image)
plt.title("Image")
plt.subplot(1,2,2)
plt.imshow(test1)
plt.title("after Intensity and Saturation selection")
Out[10]:
In [12]:
plt.figure(figsize=(40,40))
plt.subplot(1,2,1)
plt.imshow(test2)
plt.title("after gradient magtitude selection")
plt.subplot(1,2,2)
plt.imshow(test3)
plt.title("after GMM")
Out[12]:
In [13]:
plt.figure(figsize=(10,10))
# plt.subplot(1,2,1)
plt.imshow(test4)
plt.title("after fill process")
Out[13]:
第二种分割算法¶
- stage1:
- $c_E = 0.2989 · c_R + 0.5870·c_G + 0.1140·c_B$
- $r_{GE} = \frac{P_{95}(c_{G})}{P_{95}(c_E)}$ $r_{BE} = \frac{P_{95}(c_{B})}{P_{95}(c_E)}$
- $c_{G}(x_0) > r_{GE} * T_1 \mid c_{B}(x_0) > r_{BE} * T_1 \mid c_E(x_0) > T_1 $
- stage2:
- $\varepsilon(x_0) = \max \{ \frac{c_R(x_0)}{c^*_R(x_0)},\frac{c_G(x_0)}{c^*_G(x_0)},\frac{c_B(x_0)}{c^*_B(x_0)} \} $
- $c^*(x) 是一个经过填充，滤波后的图片$
- $\varepsilon(x) > T_2$
- stage3:
- $\frac{1}{N}\sum_{i=1}{N}|grad(E_n)| > T_3 \land N > N_{min}$
比较好的方法¶
- 算法要有很好的鲁棒性，不能在一张图片上过拟合
- 思想就是既要考虑召回率又要考虑准确度，加一点，减一点
- 算法要考虑到数据本身的实际情况，不能简单的拿来套用，比如要考虑消除图片上的tag，图片反光的特征
In [15]:
plt.figure(figsize=(8,8))
plt.imshow(paint_2)
plt.title("after var remove and fill")
Out[15]:
In [215]:
plt.figure(figsize=(8,8))
plt.imshow(paint_4)
plt.title("after background fillprocess")
Out[215]:
In [52]:
def show(k ) :
_, basename = os.path.split(showfile[k])
resoult = "./output/show/" + basename
plt.figure(figsize=(40,40))
plt.subplot(1,2,1)
plt.imshow(Image.open(showfile[k]))
plt.subplot(1,2,2)
plt.imshow(Image.open(resoult))
In [56]:
show(1)
In [57]:
show(2)
In [58]:
show(3)
In [79]:
show(4)
In [80]:
show(5)
In [62]:
show(9)
In [64]:
show(np.random.randint(1)%80)
In [72]:
show(np.random.randint(80))
In [78]:
show(np.random.randint(80))
